فصلنامه پردازش علائم و داده ها
سال نوزدهم شماره 3 (پیاپی 53، پاییز 1401)

Mobile social networks (MSNs) are a special type of Delay tolerant networks (DTNs) in which mobile devices communicate opportunistically to each other. One of the most challenging issues in Mobile Social Networks (MSNs) is to design an efficient message forwarding scheme that has a high performance in terms of delivery ratio, latency and communication cost. There are two different approaches for message forwarding: the single copy and the multi copy. In the single copy schemes, only one copy of a message exists in the network, which is carried by a custodian relay. On the other hand, in multi copy schemes more than one copy of the message exist. Although the multi copy schemes generally have superior performance in terms of delivery ratio and latency, they incur high congestion in the network, which consequently degrades the network performance and increases message forwarding cost. Therefore, we present a scheme for managing the number of replica and selecting appropriate relays in order to increase the performance and decrease the forwarding cost.In this paper, we present a novel message forwarding scheme, named Multi Copy Community Aware Forwarding (MCCAF), for managing the number of message replicas and selecting appropriate relays in order to increase the performance and decrease the forwarding cost. MCCAF only forwards one replica of message to each community that the destination node belongs to it. In order to find communities and nodes’ future communities, MCCAF exploits a hidden semi-Markov model (HSMM). In the proposed method, hidden states represent communities and environmental chronological information such as location, visited nodes and access points are modeled as observation in HSMM. To have a more precise community prediction, the semi-Markov model is extended to compute the likelihoods of a node being in a particular community after N transitions. After that, appropriate relays are selected to forward a message with taking into account the message expiration time, future communities of relays and the number of message replica. Evaluation results on real traces demonstrate that the MCCAF has superior performance over rival schemes in terms of delivery ratio, latency, and forwarding cost. Specifically, MCCAF yields on average about 8% better delivery ratio and about 23% less delivery delay than the nearest rival method.

Machine learning is an application of artificial intelligence that is able to automatically learn and improve from experience without being explicitly programmed. The primary assumption for most of the machine learning algorithms is that the training set (source domain) and the test set (target domain) follow from the same probability distribution. However, in most of the real-world applications, this assumption is violated since the probability distribution of the source and target domains are different. This issue is known as domain shift. Therefore, transfer learning and domain adaptation generalize the model to face target data with different distribution.In this paper, we propose a domain adaptation method referred to as IMage Alignment via KErnelized feature learning (IMAKE) in order to preserve the general and geometric information of the source and target domains. IMAKE finds a common subspace across domains to reduce the distribution discrepancy between the source and the target domains. IMAKE adapts both the geometric and the general distributions, simultaneously. Moreover, IMAKE transfers the source and target domains into a shared low dimensional subspace in an unsupervised manner.Our proposed method minimizes the marginal and conditional probability distribution differences of the source and target data via maximum mean discrepancy and manifold alignment for geometrical distribution adaptation. IMAKE maps the input data into a common latent subspace via manifold alignment as a geometric matching method. Therefore, the samples with the same class labels are collected around their means, and samples with different class are separated, as well. Moreover, IMAKE maintains the source and target domain manifolds to preserve the original data position and domain structure. Also, the use of kernels and mapping data into Hilbert space provides more accurate separation between different classes and is suitable for data with complex and unbalanced structures. The proposed method has been evaluated using a variety of benchmark visual databases with 36 experiments. The results indicate the significant improvements of the proposed method performance against other machine learning and transfer learning approaches.

The Short-term forecasting of electric load plays an important role in designing and operation of power systems. Due to the nature of the short-term electric load time series (nonlinear, non-constant, and non-seasonal), accurate prediction of the load is very challenging. In this article, a method for short-term daily and hourly load forecasting is proposed. In this method, in the first step, the intrinsic mode functions (IMFs) of the Electric load curve, which are a group of average and pseudo-periodic average signals, are extracted by using the empirical mode decomposition (EMD) method, which is a non-linear and non-constant time-frequency method. For this purpose, the maximum and minimum points of the signal are determined, and then, in one cycle, the difference between the average curve of the upper and lower envelope is calculated with it. This continues until the result falls below a threshold value, and then, the rest of the signal which contains noise is discarded to get a relatively clean signal. In the second step, we need to obtain the sub-sequences of each IMF. So, we use the wavelet transform. The wavelet transform is a kind of transform that is used to decompose a continuous signal into its frequency components, and the resolution of each component is equal to its scale. Each subsequence contains different information and details that can help the improvement of the prediction accuracy. In the third step, the obtained subsequences are aggregated and finally used for prediction by Support Vector Regression (SVR). Support vector regression is a type of supervised learning system that is used for both grouping and estimating the fitting function of data in regression problems so that the least error occurs in the grouping of data or in the fitting function. The purpose of the proposed method is to reduce the error for daily and hourly load prediction. In this method, two datasets of Poland and Canada have been experimented. With four criteria of mean square error (MSE), root mean square error (RMSE), average absolute percentage error (MAPE) and mean absolute error (MAE), the results are evaluated. The findings show that the load prediction error for the Polish data set are as follows: MSE equal to 0.0012, RMSE equal to 0.0342, MAPE equal to 2.9771, and MAE equal to 0.0044. For Canadian data set, the results are as follows: MSE equal to 5.0969e-07, RMSE equal to 7.1393e-04, MAPE criterion equal to 0.9571, and the MAE criterion equal to 2624e-04. Comparison of the proposed method with other competing methods show that better results are achieved by the proposed method in term of the error rate.

Addiction is a biological, psychological, and social disease. Several factors are involved in etiology, substance abuse, and addiction which interact with each other and lead to the beginning of drug use and then addiction. Heroin is an addictive drug that, by acting on the central nervous system, reduces the density of neurons in the brain and interferes with decision making. This paper examines the effects of heroin on brain function by studying the relationship between spectral strength of electroencephalogram (EEG) signal and heroin abuse. For this purpose, the resting EEG signal and cognitive activity of 15 healthy individuals and 15 heroin addicts were recorded in 16 channels in one session. The frequency range of EEG signal sub-bands was calculated separately for each individual. Welch method has been used to extract the power of EEG signal frequency sub-bands. The extracted features were examined using Mann-Whitney test and Davies-Bouldin index. The results show that the heroin-dependent group has higher power in delta (in the frontal, central and temporal regions) and theta (in all canals) than in the control group. In the heroin-dependent group, the power of alpha decreased compared to the control group. High alpha sub-bands power in the frontal, temporal and central lobes compared to other frequency sub-bands, as well as in the central, parietal and temporal lobes, the power of the second low alpha sub-band in decreased addicts. According to Davies-Bouldin, the power of the second low alpha sub-band in the T6 channel has a better power to differentiate between healthy and heroin-dependent people.

The use of artificial intelligence in the process of diagnosing heart disease has been considered by researchers for many years. In this paper, an efficient method for selecting appropriate features extracted from electrocardiogram (ECG) signals, based on a genetic algorithm for use in an ensemble multi-kernel support vector machine classifiers, each of which is based on an optimized genetic algorithm is proposed. It has already been shown that due to its features (feature space mapping and decision boundary maximization), support vector machine classification is one of the classification methods that are suitable for any type of environment. This paper uses a number of multi-kernel support vector machine classifiers as an ensemble classifier. ensemble diversity is created by teaching each multi-kernel support vector machine classifier on a subspace (ie, a subset of features). In this method, the majority vote method is used to combine the output of the categories. On the other hand, in the classification of ECG signals, signals are usually used as their characteristics; As a result, since the methods of classifying signals are faced with a large number of features, and not removing these features creates a problem of high dimensions and also increases the computational for the intended application, the step of selecting the feature is inevitable. The extracted features include temporal properties, AR, and wavelet coefficients, the number of which will be optimized using a genetic algorithm. The evaluation of this set of features selected by the genetic algorithm is examined by applying it to a multivariate SVM. A genetic algorithm is used to optimize the parameters of each of the SVMs. Indicates the desired method. With the help of computer simulation, the overall accuracy of the system for identifying 6 types of heart rhythms is 99.15%, which in comparison with the accuracy obtained with previous research, shows the optimal performance of the proposed method.

With the advancement of technology, the use of ATM and credit cards are increased. Cyber fraud and theft are the kinds of threat which result in using these Technologies. It is therefore inevitable to use fraud detection algorithms to prevent fraudulent use of bank cards. Credit card fraud can be thought of as a form of identity theft that consists of an unauthorized access to another person's card information for the purpose of charging purchases to the account or removing funds from it. Credit card fraud schemes are divided into two categories: application fraud and account takeover. When a credit card account gets opened without someone’s permission is called application fraud. Account takeovers, on the other hand, is when an existing credit card account is hijacked, and the criminal obtains enough personal information to modify the account's information. The criminal then subsequently reports the card lost or stolen in order to obtain a new card and make unauthorized purchases with it. Data mining as a technique capable of identifying useful patterns among a great deal of data is an effective method in detecting fraud in this regard. The main purpose of this paper is to present a new method for detecting unattended outliers that require high accuracy and recall. The method presented in this study is based on a combination of NMF, hierarchical k-means, k-means and k-nearest neighbors’ techniques. To evaluate the proposed method of outlier detection, several experiments were performed using standard data, in terms of accuracy and recall with Isolation Forest, k-nearest neighbors, Median kNN, and Average kNN. The dataset used in this paper is one that was provided in a 2016 Kaggle competition and was provided by a European bank after anonymization. The results, corroborate that the proposed method has higher accuracy and recall than other algorithms.

Linear discriminant analysis is a well-known matrix-based dimensionality reduction method. It is a supervised feature extraction method used in two-class classification problems. However, it is incapable of dealing with data in which classes have unequal covariance matrices. Taking this issue, the Chernoff distance is an appropriate criterion to measure distances between distributions. In the proposed method, for data classification, LDA is used to extract most discriminative features but instead of its Fisher criterion, the Chernoff distance is employed to preserve the discriminatory information for the several classes with heteroscedastic data. However, the Chernoff distance cannot handle the situations where the component means of distributions are close and leads to the component distribution overlap and underperforming classification. To overcome this issue, the proposed method designs an instance selection method that provides the appropriate covariance matrices. Aiming to improve LDA-based feature selection, the proposed method includes two phases: (1) it removes non-border instances and keeps border ones by introducing a maximum margin sampling method. The basic idea of this phase is based on keeping the hyperplane that separates a two-class data and provides large margin separation. In this way, the most representative instances are selected. (2) It extracts features on selected instances by the proposed extension of LDA which generates a desirable scatter matrix to increase the efficiency of LDA. In the proposed method, the instance selection process is considered a constrained binary optimization problem with two contradicting objects, and the problem solutions are obtained by using a heuristic method named filled function. This optimization method does not easily get stuck in local minima; meanwhile, it is not affected by improper initial points. The performance of the proposed method on data collected from the UCI database is evaluated by 10-fold validation. The results of experiments are compared to several competing methods, which show the superiority of the proposed method in terms of classification accuracy percentage and computational time.

The Iranian traditional music is composed of seven majors Dastgahs: Chahargah, Homayoun, Mahour, Segah, Shour, Nava, and Rast-Panjgah. In this paper, a new algorithm for the classification of the Iranian traditional music Dastgahs based on pitch frequency is proposed. In this algorithm, the features of Lagrange coefficients of pitch logarithm (LCPL), Fuzzy similarity sets type 2 (FSST2), and their combination are used as the representation of music signals which are fed into the multi-class support vector machine (MSVM) as the classifier. The features of LCPL and FSST2 are obtained by applying some modifications on the pitch frequency of the desired music. To compute LCPL, first, the values of pitch frequency are extracted by the PRAAT algorithm. Then, after the applying a logarithmic operation, the tracks of pitch frequency are partitioned into smaller segments. The method of feature extraction is based on detecting the trough or valley points of the pitch tracks. In the following, the coordinates of trough points (i.e., the index of pitch frequency and the logarithmic value of the frequency) are considered as each segment boundaries. In the next step, the track between the two boundaries of each segment is approximated by a 6th order Lagrange polynomial and the computed polynomial coefficients are considered as a 6-dimensional feature vector. The first step in extracting the FSST2 feature is to compute the pitch frequencies of the input signal by the PRAAT algorithm. The second step involves the classification of music notes. Then, the subtractive clustering method is used to eliminate the incorrectly estimated pitch frequencies of the previous step. Next, the process of folding notes (i.e., transferring the extracted pitch frequencies into the reference octave band of 220-440 Hz) is performed followed by translating the frequency points to the cents with respect to 220 Hz. After folding notes in one octave, the Mahalanobis distance is applied to recognize which point on the reference octave corresponds to each musical note. These same procedures are conducted for the information pattern (theoretical data) of each Dastgah. In the final step, the folded frequency points of the unknown input signal and the information pattern of all Dastgahs are transferred to the Fuzzy logicType-2 domain and compared to determine a similarity measure which is considered as the extracted feature. The dataset used in the proposed classification algorithm contains the excerpts from solo performances with Tar played by Alizadeh, the well-known Iranian music master. The performances of the baselines and proposed classification algorithms are evaluated by the measures of Accuracy, Recall, Precision, F-measure, and MCC. The results show that the proposed algorithm has a better performance as compared with the baseline methods in terms of different classification criteria.

Today, the use of wireless and intelligent networks are widely used in many fields such as information technology and networking. There are several types of these networks that MPLS networks are one of these types. However, in MPLS networks there are issues and problems in the design and implementation discussion, for example security, throughput, losses, power consumption and so on. Basically, the performance parameters of MPLS networks are directly dependent on the type of routing performed between these networks. In this paper, a routing algorithm for energy-based routing in MPLS networks is presented which performs routing online with increasing number of requests. This algorithm guarantees bandwidth and end-to-end delay and its performance is based on the Minimum Delay and Maximum Flow (MDMF) algorithm. The purpose of this algorithm is to uniformly distribute the load across the network for all requests and to delay high-bandwidth requests and maximum end-to-end delay using a fuzzy law-based approach. In addition, the proposed algorithm optimizes the use of resources and increases the number of requests accepted. In the proposed routing algorithm, the weight of each link of the network graph is first calculated based on a multi-objective weighted sum maximization problem. The weight of links must be calculated in such a way as to determine the shortest path weighted by the sum of the minimum weights in the graph. Then, links which their residual bandwidth is less than the requested bandwidth are removed from the grid graph and in the residual subgraph, the shortest weighted path between the input-output nodes is calculated using Dijkstra algorithm and based on the weights of each link. Therefore, this algorithm only requires the remaining capacity information of the links, where assumed this information is always available. The proposed method is implemented in MATLAB software and the results show a decrease in energy consumption and improvement of network performance parameters compared to similar methods.

K-nearest neighbors (KNN) based recommender systems (KRS) are among the most successful recent available recommender systems. These methods involve in predicting the rating of an item based on the mean of ratings given to similar items, with the similarity defined by considering the mean rating given to each item as its feature. This paper presents a KRS developed by combining the following approaches: (a) Using the mean and variance of item ratings as item features to find similar items in an item-wise KRS (IKRS); (b) Using the mean and variance of user ratings as user features to find similar users with a user-wise KRS (UKRS); (c) Using the weighted mean to integrate the ratings of neighboring users/items; (d) Using ensemble learning. Three proposed methods EVMBR, EWVMBR and EWVMBR-G are presented in this paper. All three methods are user-based, in which VM distance is used as a measure of the difference between users / items, to find neighboring users / items, and then the weighted average is weighted, respectively. Also, weights based on the Gaussian combined covariance model are used to predict unknown user ratings. Our empirical evaluations show that the proposed method EVMBR, EWVMBR and EWVMBR-G, which utilizes ensemble learning, are the most accurate among the methods evaluated. Depending on the dataset, the proposed method EWVMBR-G managed to achieve 20 to 30 percent lower mean absolute error than the original MBR. In terms of runtime, the proposed methods are comparable to the MBR and much faster than the slope-one method and the cosine- or Pearson-based KNN recommenders.

Timely diagnosis of fatigue helps to improve the quality and effectiveness of neurofeedback training. Neurofeed back training (NFT) is a method that can change brain activity by altering brain signal fluctuations and teaches individuals to produce or reproduce their brain activity patterns in order to improve performance. Neurofeedback training has been widely utilized over the recent years owing to its considerable effect on the cognitive processes. Fatigue during NFT is one of factors affecting the functioning and achievement of NFT which results in decreased learning ability. Timely diagnosis of fatigue during NFT preserves quality of NFT.Decreased learning ability reduces individuals' motivation for learning during NFT. In this paper, 12 participants` electroencephalogram signals were investigated to detect fatigue during NFT. Two training protocols named protocol 1 and 2 have been designed to improve working memory. Each protocol includes 6 participants and 10 training sessions that each session takes three 10-minute training intervals. Training features in protocol 1 are increased in power of lower2 alpha frequency band in OZ channel and permutation entropy reduction in FZ channel, while protocol 2`s training feature is increased in power of lower2 alpha frequency band in OZ channel. Occurrence of fatigue during NFT changes trend of training features. Changing of training features slope will decrease or become opposite to the goal of NFT. Therefore, examining trend of training features slope is a novel approach in detection of fatigue during NFT. During the occurrence of fatigue, in protocol 1, trend of power of lower2 alpha frequency band`s slope in the OZ channel is decreasing and the trend of entropy` s slope in the FZ channel is increasing. Consequently, the trend of score`s slope is also decreasing. Also in the protocol 2, the trend of power of lower 2 alpha frequency band`s slope in the OZ channel and score is decreasing. This shows that training features do not change in line with the neurofeedback`s goal. Fatigue was detected for 3 subjects in the protocol 1 and 1 subject in the protocol 2. Occurrence of fatigue was less in protocol 1 compared with protocol 2 since Protocol 1 `s training features are combination of frequency and non-frequency features, while the Protocol 2 `s training feature is only frequency feature. Detection of fatigue during NFT is an essential issue which contributes to increase in the effect of training and participants` performance.

The main task of the tokenization is to divide the sentences of the text into its constituent units and remove punctuation marks (dots, commas, etc.). Each unit is a continuous lexical or grammatical writing chain that is an independent semantic unit. Tokenization occurs at the word level and the extracted units can be used as input to other components such as stemmer. The requirement to create this tool is to identify and recognize the units that are known as independent semantic units in Persianlanguage. This tool detects word boundaries in texts and converts the text into a sequence of words.In the English language, many activities have been done in the field of text tokenization and many tools have been development; such as: Stanford, Ragel, ANTLR, JFLex, JLex, Flex and Quex. In recent decades, valuable researches have also been conducted in the field of tokenization in Persian language that all of them have worked on the lexical and syntactic layer. In the current research, we tried to focus on the semantic layer in addition to those two layers.Persian texts usually have two simple but important problems. The first problem is multi-word tokens that result from connecting one word to the next. Another problem is polysyllabic units, which result from the separation of words that together form a lexical unit.  Tokenizer is one of the language preprocessing tools that is widely used in text analysis. This component recognizes the center of words in texts and turns it into a sequence of words for later analysis. Variety in Persian script and non-observance of the rules of separation and spelling of words on the one hand and the lexical complexities of Persian language on the other hand, different language processing such as tokenization face many challenges. Therefore, in order to obtain the optimal performance of this tool, it is necessary to first specify the computational linguistics considerations of tokenization in Persian and then, based on these considerations, provide a data set for training and testing. In this article, while explaining the mentioned considerations, we tried to prepare a data set in this regard. The prepared data set contains 21.183 tokens and the average length of sentences is 40.28.